GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ProprietaryLegal
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
ProprietaryLegal
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
We use Qwen3.8-27B for English-language legal drafting. In that setting the CJK part of the vocabulary does no useful work, and sometimes does harm: the base model occasionally emits a stray Chinese character mid-sentence in English output. Removing those entries:
It is a subtractive, verifiable change, so it can be proven not to alter English behaviour.
An entry was removed when any text containing its bytes must contain a CJK code point. That covers:
| Removed | Entries |
|---|---|
| Han (Chinese characters, incl. Japanese kanji) | 54,715 |
| Hangul (Korean) | 6,807 |
| Hiragana | 2,487 |
| Katakana | 1,678 |
| CJK and fullwidth punctuation/forms | ~250 |
| Byte fragments that can only complete to CJK | 601 |
| Total | 66,533 |
The CJK script set covers Hangul Jamo, CJK radicals, ideographic description characters, CJK symbols and punctuation, kana, Bopomofo, enclosed CJK letters, CJK compatibility, CJK Unified Ideographs with extensions A–H, compatibility ideographs and forms, and halfwidth/fullwidth forms.
Kept, deliberately:
§, ¶, curly quotes, en/em dashes, accented Latin letters, Greek, math symbols and digits. ASCII-only trims lose these, and they matter in legal text;A broader trim that also removed the other non-Latin scripts would have dropped 103,447 entries (2.12 GB) and changed the tokenization of none of our English test documents. We chose not to do it, so the model keeps its other languages.
Dependency closure: the tokenizer is byte-level BPE, so a kept entry can depend on a merge that passes through a removed one. We checked every retained merge. None needed a removed entry: 0 entries had to be restored and there were 0 closure errors. The merge list is rebuilt without the removed merges, and no library helper that renumbers IDs was used.
All work ran on one NVIDIA DGX Spark (GB10), using torch 2.11.0+cu130, transformers 5.18.0 and tokenizers 0.23.2. The base is Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. All 19 files were checked against their SHA-256 hashes before and after; the base was only read.
tokenizer.json, vocab.json, merges.txt and tokenizer_config.json (including added_tokens_decoder) were rewritten.remap/id-remap.json holds the full new_to_old map for all 181,760 rows. The 216 padding rows are the base's own padding rows.model.language_model.embed_tokens.weight and lm_head.weight (both [181760, 5120]; the tables are untied) were rewritten with the kept rows in new-ID order.The surgery, export and proof code is in trim-code/ (tokenizer_surgery.py, export_trim.py, prove_trim.py). provenance.json records the base revision, the tokenizer hash and the SHA-256 of every shard.
The test text was a private corpus of about 3,500 English legal documents: pleadings, motions, affidavits, orders, agreements and correspondence, plus treatise and statute text. It is not released.
| Check | Result |
|---|---|
| Tokenization of English legal text | 3,526 / 3,526 documents tokenize identically after the remap |
| Byte-exact decode round trips | 3,540 / 3,540 |
| Logits vs. base (32 held-out excerpts, ≤384 tokens each) | bit-identical: max abs difference 0.0, KL 0.0 |
| Greedy generations vs. base (24 English prompts) | 24 / 24 token-for-token identical |
| CJK characters in English output | 0 |
| Chinese-language prompts | Base answered in Chinese; the trimmed model answered in English with 0 CJK characters |
§, ¶, curly quotes, dashes, accented names |
The logits are bit-identical because the removed rows only ever competed in the output softmax. On text where the base never chose them, every remaining logit is computed by exactly the same weights.
The interface is the same as the base model's:
python
from transformers import AutoProcessor, AutoModelForImageTextToText model_id = "ProprietaryLegal/Qwen3.8-27B-CJK-Trim"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="bfloat16", device_map="auto") messages = [{"role": "user", "content": [{"type": "text", "text": "Explain what a pendente lite order is."}]}]inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512)print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))We tested it with transformers 5.18. vLLM, llama.cpp and other runtimes should work, because the architecture is unchanged and only vocab_size differs, but we have not tested them yet.
remap/id-remap.json.embed_tokens or lm_head apply unchanged, because every other weight is identical. Adapters that do touch them need their rows remapped.generation_config, added_tokens, image ID, vocab.json, merges.txt); none is present here.Apache 2.0, inherited from Qwen/Qwen3.8-27B. The original LICENSE file is included unchanged. All credit for the model itself belongs to the Qwen team.
vocab_sizeconfig.jsongeneration_config.jsonAutoProcessor loads, and the processor's image and video token IDs are correct.| Intact |
| Base model's probability mass on the removed entries, on held-out English legal text | mean 0.03%, max 5.7% (now unavailable) |
| Load time / memory on DGX Spark | 70.9 s / 50.9 GB (base 74.3 s / 52.2 GB) |